AI & Machine Learning Roadmap

Master data processing, modeling, and deep learning algorithms.

Overall Progress0%

Unsupervised Learning

Explanation

Unsupervised learning finds structure in data that has no labels. Instead of learning "this input maps to this answer," the algorithm discovers patterns on its own: groups of similar items, reduced representations, or unusual points that do not fit. This lesson focuses on clustering — especially K-means, which you will work through step by step — plus dimensionality reduction (PCA), anomaly detection, and the real-world uses like customer segmentation. You will also learn when to choose unsupervised over supervised methods. The examples use small, manually understandable datasets so you can follow the algorithm without needing external libraries.

Labeled vs Unlabeled Data

	Labeled	Unlabeled
Has a target?	Yes	No
Example	Each email marked spam/not	Raw customer records, no category
Goal	Predict the label	Find hidden structure

Clustering and Similarity

Clustering groups similar items together.
Similarity is measured by distance in feature space.
Items in the same cluster are close; items in different clusters are far apart.

Distance and Feature Space

Each data point is a position whose coordinates are its features.
Distance (e.g., Euclidean) measures how far apart two points are.
The choice of distance and feature scale strongly affects clusters.

K-means Clustering

K-means partitions data into k clusters. Each cluster has a centroid (its center). The algorithm alternates between assigning points to the nearest centroid and moving centroids to the mean of their assigned points.

Choose k (the number of clusters).
Place k initial centroids.
Assign each point to the nearest centroid.
Move each centroid to the mean of its points.
Repeat assignment and update until stable.

Choosing k

k is chosen by the analyst, not learned automatically.
The elbow method plots error versus k and looks for a bend.
Domain knowledge (e.g., "we have three customer segments") also guides k.

Centroids and Assignment

A centroid is the center (mean) of its cluster.
Assignment picks the closest centroid for each point.
Because centroids move, assignment can change between iterations.

Hierarchical Clustering (Concept)

Builds a tree of clusters by merging the closest ones (or splitting).
Produces a hierarchy rather than a flat set of k clusters.
Useful when the number of clusters is unknown.

Dimensionality Reduction and PCA (Concept)

Dimensionality reduction compresses many features into fewer.
PCA finds the directions that capture the most variance.
Reduced data is easier to visualize and can speed up other models.

Anomaly Detection (Concept)

Find points that are unusual relative to the rest.
Often framed as points far from any cluster or outside a normal region.
Used for fraud, faults, and intrusion detection.

Real-World Uses

Market/customer segmentation: group users by behavior.
Recommendation and discovery: find related items.
Anomaly detection: flag suspicious transactions.
Data exploration: understand structure before labeling.

Strengths and Limitations

Strength: works without labels, reveals hidden structure.
Limitation: results are hard to evaluate objectively (no "correct" answer).
K-means assumes roughly round, similar-size clusters and requires choosing k.

Supervised vs Unsupervised

	Supervised	Unsupervised
Input	Labeled data	Unlabeled data
Output	Predicted label/value	Clusters / structure
Evaluation	Compare to true labels	Harder, more subjective

Common Mistakes

Ignoring feature scale before computing distances.
Treating clustering output as ground truth.
Choosing k without any reasoning.
Interpreting clusters as meaningful groups without validation.

Examples

K-means on a small 1D dataset

Grouping numbers into two clusters.

text

1Points: 1, 2, 3, 8, 9, 10  (k = 2)2Start centroids at 1 and 83Assign: {1,2,3} and {8,9,10}4Update: centroids → 2 and 95Stable: clusters stay {1,2,3} and {8,9,10}

Output

(two natural groups emerge)

Customer segmentation

Grouping shoppers by behavior.

text

1Features: monthly spend, visit frequency2Clusters found:3  A: low spend, low visits (casual)4  B: high spend, high visits (loyal)5  C: high spend, low visits (big-ticket)

Output

(segments inform marketing strategy)

Key Takeaways

Unsupervised learning finds structure without labels.
Clustering groups similar items by distance in feature space.
K-means alternates assignment and centroid update until stable.
Choose k deliberately; scale features before measuring distance.
Dimensionality reduction (PCA) compresses data; anomaly detection flags outliers.
Unsupervised results are exploratory, not ground truth.

← PreviousSupervised Learning Next →Model Evaluation